▶ 70:09 · World Models & Simulation · World action models with latent future prediction surpass VLAs but require architectural breakthroughs for efficiency

episode briefing

Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club

2026-08-08 · 8 company · 11 thematic

sentiment

7 bull1 bear0 neu

speakers

nico

Co-founder of Default (default.com), providing real-time data layer and semantic modeling for B2B companies to power AI agents. Raised Series A. Previously built inbound scheduling/routing tools for fast-growing startups. Company has logo tattoo culture.

bill

Technical background in VLMs and foundation models (worked at Siemens on time series prediction). Co-founded General Instinct to build infrastructure for running physical AI models (world action models, VLAs) fast on edge devices.

guanming

Robotics RL background. Co-founded General Instinct with Bill to optimize world action models for real-time deployment, achieving 500ms per chunk on Jetson via distillation, architecture changes, and modality innovations.

marcel

Co-host of itnig live tertulias; focuses on go-to-market strategy and AI adoption in enterprise software; expresses FOMO about speed of AI implementation in sales/marketing.

milan ganai

PhD student with Marco Pavone and Clark Barrett at Stanford, currently working at Waymo. Presented RNB encore, a self-supervised bootstrapping framework for discovering embodiment-specific reasoning across manipulation, locomotion, and driving.

now playing · World Models & Simulation

Robotics & Physical AItailwindscore 8/10marcel

Hierarchical memory (short-term visual + long-term language) enables long-horizon robot tasks and in-context adaptation

Marcel's MEM system shows that decomposing memory into dense short-context frames for dexterity and compressed language long-context for task progress allows VLAs to execute 20+ minute task…

Robotics & Physical AItailwindscore 8/10marcel

Memory in robotics policies enables long-horizon tasks and in-context adaptation

Decomposing memory into short-term dense visual frames for dexterity and long-term compressed language representations for task progress allows policies to handle 10+ minute tasks, wait app…

AI Agentstailwindscore 7/10milan ganai

Embodied reasoning must be discovered per-embodiment via self-supervised bootstrapping, not hand-designed

Milan's RNB encore uses variational inference to automatically discover which reasoning types (move+gripper for manipulation, structural affordances for locomotion, meta-action for driving)…

Robotics & Physical AItailwindscore 7/10milan ganai

Self-supervised bootstrapping discovers embodiment-specific reasoning for data-scarce robotics

Treating reasoning as a latent variable and using variational inference to score reasoning traces for concision, non-triviality, and action predictiveness automatically discovers optimal re…

World Models & Simulationtailwindscore 8/10tyler lum

Sim-to-real RL with goal-conditioned policies achieves zero-shot dexterous tool use on real robots

Tyler demonstrates that training a single goal-conditioned policy in GPU-accelerated simulation on primitive objects with random goals transfers zero-shot to novel real-world tools and task…

Robotics & Physical AItailwindscore 9/10nico

Robotics application companies emerging as the new SaaS wave for physical world automation

End-to-end robotics application companies (data centers, warehouses, manufacturing, food) that start with teleop on off-the-shelf hardware, iterate fast with real customers, and build opera…

Robotics & Physical AItailwindscore 8/10nico

Robotics application companies emerging as the new SaaS for the physical world

Nico argues that full-stack robotics application companies (teleop-first, off-the-shelf hardware, fine-tuning models) will transform the physical economy like SaaS did for software, enabled…

AI Infrastructuretailwindscore 7/10nico

Specialized data infrastructure required for physical AI's multimodal, multi-rate, episodic 3D data

Nico explains that physical robotics data (multimodal, multi-rate, episodic, 3D semantics, deep nested structures) breaks traditional databases, creating a need for purpose-built storage an…

AI Infrastructuretailwindscore 8/10nico

Physical AI data layer and inference optimization critical for robotics deployment at scale

Robotics data is fundamentally different from web data (multimodal, multi-rate, episodic, 3D semantics) requiring specialized storage/query tools (Rerun), while world action models need 50x…

AI Infrastructuretailwindscore 7/10bill

World action models need inference optimization infrastructure to run on edge devices

Bill and Guanming show that raw world action models are prohibitively slow (200s, $70K) for robotics, but distillation of VAEs and DiTs, cross-attention between video and action transformer…

World Models & Simulationtailwindscore 8/10guanming

World action models with latent future prediction surpass VLAs but require architectural breakthroughs for efficiency

World action models (Dream Zero, General Instinct) that predict future kinematics via diffusion transformers achieve better physics understanding than VLAs, but current implementations are…

←

▶ 70:09 · World Models & Simulation · World action models with latent future prediction surpass VLAs but require architectural breakthroughs for efficiency

episode briefing

Y Combinator

Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club

2026-08-08 · 8 company · 11 thematic

sentiment

7 bull1 bear0 neu

speakers

Co-host of itnig live tertulias; focuses on go-to-market strategy and AI adoption in enterprise software; expresses FOMO about speed of AI implementation in sales/marketing.

milan ganai

now playing · World Models & Simulation

Robotics & Physical AItailwindscore 8/10marcel

Hierarchical memory (short-term visual + long-term language) enables long-horizon robot tasks and in-context adaptation

Marcel's MEM system shows that decomposing memory into dense short-context frames for dexterity and compressed language long-context for task progress allows VLAs to execute 20+ minute task…

Robotics & Physical AItailwindscore 8/10marcel

Memory in robotics policies enables long-horizon tasks and in-context adaptation

Decomposing memory into short-term dense visual frames for dexterity and long-term compressed language representations for task progress allows policies to handle 10+ minute tasks, wait app…

AI Agentstailwindscore 7/10milan ganai

Embodied reasoning must be discovered per-embodiment via self-supervised bootstrapping, not hand-designed

Milan's RNB encore uses variational inference to automatically discover which reasoning types (move+gripper for manipulation, structural affordances for locomotion, meta-action for driving)…

Robotics & Physical AItailwindscore 7/10milan ganai

Self-supervised bootstrapping discovers embodiment-specific reasoning for data-scarce robotics

Treating reasoning as a latent variable and using variational inference to score reasoning traces for concision, non-triviality, and action predictiveness automatically discovers optimal re…

World Models & Simulationtailwindscore 8/10tyler lum

Sim-to-real RL with goal-conditioned policies achieves zero-shot dexterous tool use on real robots

Tyler demonstrates that training a single goal-conditioned policy in GPU-accelerated simulation on primitive objects with random goals transfers zero-shot to novel real-world tools and task…

Robotics & Physical AItailwindscore 9/10nico

Robotics application companies emerging as the new SaaS wave for physical world automation

End-to-end robotics application companies (data centers, warehouses, manufacturing, food) that start with teleop on off-the-shelf hardware, iterate fast with real customers, and build opera…

Robotics & Physical AItailwindscore 8/10nico

Robotics application companies emerging as the new SaaS for the physical world

Nico argues that full-stack robotics application companies (teleop-first, off-the-shelf hardware, fine-tuning models) will transform the physical economy like SaaS did for software, enabled…

AI Infrastructuretailwindscore 7/10nico

Specialized data infrastructure required for physical AI's multimodal, multi-rate, episodic 3D data

Nico explains that physical robotics data (multimodal, multi-rate, episodic, 3D semantics, deep nested structures) breaks traditional databases, creating a need for purpose-built storage an…

AI Infrastructuretailwindscore 8/10nico

Physical AI data layer and inference optimization critical for robotics deployment at scale

Robotics data is fundamentally different from web data (multimodal, multi-rate, episodic, 3D semantics) requiring specialized storage/query tools (Rerun), while world action models need 50x…

AI Infrastructuretailwindscore 7/10bill

World action models need inference optimization infrastructure to run on edge devices

Bill and Guanming show that raw world action models are prohibitively slow (200s, $70K) for robotics, but distillation of VAEs and DiTs, cross-attention between video and action transformer…

World Models & Simulationtailwindscore 8/10guanming

World action models with latent future prediction surpass VLAs but require architectural breakthroughs for efficiency

World action models (Dream Zero, General Instinct) that predict future kinematics via diffusion transformers achieve better physics understanding than VLAs, but current implementations are…